From b2a9d32a09e6afb3a3fbe7d90df1ae96e54e1377 Mon Sep 17 00:00:00 2001 From: Claude Date: Thu, 26 Mar 2026 18:42:39 +0000 Subject: [PATCH] Pivot from local AI rig to media server: add R730xd comparison for Blu-ray ripping MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Local AI GPU build deferred — $1,000 in GPUs can't match Claude Max for complex multi-file coding tasks. Added server comparison (E5-2620/2660/2690 V4), Blu-ray/DVD ripping requirements (MakeMKV, HandBrake, USB expansion), RAID controller analysis, and recommendation for E5-2660 V4 + H730 as best media server config. Preserved all prior GPU research for future reference. https://claude.ai/code/session_01PtYTPherSJaxDEVPgF6Nxu --- docs/gpu-setup-research.md | 187 ++++++++++++++++++++++++++++++++++++- 1 file changed, 186 insertions(+), 1 deletion(-) diff --git a/docs/gpu-setup-research.md b/docs/gpu-setup-research.md index 9519d87..8dd3b82 100644 --- a/docs/gpu-setup-research.md +++ b/docs/gpu-setup-research.md @@ -1,6 +1,6 @@ # GPU Setup Research: Rack Server AI Workloads -*Last updated: March 22, 2026* +*Last updated: March 26, 2026* ## Goal Cost-efficient rack-mountable GPU setup for: @@ -846,6 +846,182 @@ INVOKEAI_PRECISION=float16 # ComfyUI — no special flags needed ``` +--- + +## Decision: Local AI Deferred — Claude Max + Media Server (March 26, 2026) + +### Why the GPU Rig Is Off the Table (For Now) + +After thorough analysis, the $1,000 dual RTX 5000 build doesn't make economic sense +for the primary use case: **complex multi-file coding on 10,000+ line codebases**. + +**The core problem:** Tasks like "add encrypted sync across a web extension" require +frontier-level reasoning (Claude Opus) + massive context (200K+ tokens). The best local +model at 32GB VRAM (Qwen 2.5 Coder 32B) scores ~55% on SWE-Bench vs Claude Opus at 80.9%. +That gap is the difference between "gets it done" and "burns hours failing." + +| Factor | Local (2x RTX 5000) | Claude Max | +|--------|---------------------|------------| +| Cost | ~$1,000 once | $100/mo ($1,000 = 10 months) | +| Complex coding quality | A- (SWE-Bench ~55-73%) | A++ (SWE-Bench 80.9%) | +| Context window | ~32K-128K tokens | 200K tokens | +| 10K-line codebase | Partial, chunked | Entire project in one shot | +| Rate limits | None | None (Max tier) | +| Hardware depreciation | Loses value over time | Always current model | +| Electricity | ~$9/mo | $0 | + +**The catch-22:** AI demand is driving up hardware costs. DDR4 ECC RDIMM prices jumped +~90% QoQ in Q1 2026. Used GPU prices are climbing, not falling. The old "wait 6 months, +get 2x for the same price" contract is broken — memory fabs prioritize HBM for data centers +over consumer/used server markets. + +**The plan:** Subscribe to Claude Max ($100/mo), revisit local AI when open-source models +close the gap on complex architectural coding tasks (estimated 1-2 years). The $1,000 +buys dramatically more local capability in 2027-2028. + +### New Server Purpose: Media Server + General Homelab + +With GPUs off the table, the R730xd becomes a **media/storage server** with these goals: +1. **Blu-ray/DVD ripping** — Multiple external ODD drives simultaneously via MakeMKV +2. **HandBrake encoding** — CPU-heavy transcoding (the Xeon cores actually matter now) +3. **Media serving** — Plex/Jellyfin +4. **General homelab** — Docker, storage, VMs, network services + +--- + +## R730xd Server Comparison: Media Server Build + +### The Three Candidates + +| Spec | **E5-2620 V4** | **E5-2660 V4** | **E5-2690 V4** | +|------|----------------|----------------|----------------| +| **Cores/Threads** | 8x2 = 16C/32T | 14x2 = 28C/56T | 14x2 = 28C/56T | +| **Base/Turbo** | 2.1 / 2.5 GHz | 2.0 / 2.4 GHz | 2.6 / 3.5 GHz | +| **TDP per CPU** | 85W | 105W | 135W | +| **Total CPU TDP** | 170W | 210W | 270W | +| **RAID** | H330 | H730 | H730P | +| **RAM** | 32GB | 32GB | 32GB | +| **Drives included** | 2x trays | 2x trays | 2x trays | + +### What Matters for Blu-ray/DVD Ripping + +**MakeMKV (ripping)** is almost entirely **I/O-bound**, not CPU-bound. Even an old Celeron +can rip 2 discs at once at single-digit CPU usage. MakeMKV just decrypts and copies — the +bottleneck is the optical drive read speed and disk write speed. 16 cores is overkill for +ripping alone. + +**HandBrake (encoding)** is **heavily CPU-bound**. This is where cores and clock speed matter. +Encoding a single Blu-ray with x265 can peg all cores at 100% for 30-60+ minutes. Multiple +simultaneous encodes scale linearly with available cores. + +| Task | CPU Need | Bottleneck | +|------|----------|------------| +| MakeMKV rip (per drive) | ~2-5% of 1 core | Optical drive read speed | +| 3x simultaneous MakeMKV rips | ~5-15% total | Drive speed + disk write | +| HandBrake x265 encode (1080p) | 100% all cores | Pure CPU | +| HandBrake + 3x MakeMKV rips | 90%+ CPU for HB, trivial for MKV | CPU for encode | +| Plex transcoding (1 stream) | ~2-4 cores at moderate load | CPU or iGPU | + +### RAID Controller Matters + +| Controller | Cache | RAID Levels | Battery | Notes | +|------------|-------|-------------|---------|-------| +| **H330** | None | 0,1,5,10,50 | No | Basic HBA-ish. No write cache. Fine for JBOD/passthrough | +| **H730** | 1GB | 0,1,5,6,10,50,60 | Yes (BBU) | Write cache = faster writes. Good for ripping to RAID | +| **H730P** | 2GB | 0,1,5,6,10,50,60 | Yes (BBU) | Best controller. 2GB cache handles burst writes well | + +For ripping multiple discs simultaneously to spinning drives, the **write cache matters**. +Each Blu-ray drive reads at ~35-55 MB/s. Three drives = ~105-165 MB/s sustained writes. +The H730/H730P cache absorbs bursts; the H330 writes directly, which can bottleneck on +spinning disks. + +### USB Connectivity for External ODD Drives + +The R730xd has **very few USB ports**: + +| Location | Quantity | Standard | +|----------|----------|----------| +| Front | 1 | USB 2.0 (iDRAC management) | +| **Rear** | **2** | **USB 3.0** | +| Internal | 1 | USB 3.0 | + +**2 rear USB 3.0 ports is not enough for multiple drives.** You need a PCIe USB 3.0 card. + +#### Recommended: PCIe USB 3.0 Expansion + +| Card | Ports | Power | Notes | +|------|-------|-------|-------| +| **StarTech PEXUSB3S44V** | 4x USB 3.0 | SATA/LP4 (optional) | 4 dedicated 5 Gbps channels — each port gets full bandwidth | +| HighPoint RocketU 1344A | 4x USB 3.0 | PCIe slot only | Confirmed working in Dell PowerEdge servers | +| SIIG 7-Port USB 3.0 | 7x USB 3.0 | Needs aux power | Overkill but available | + +**Important:** The R730xd uses **low-profile PCIe brackets** in the risers. Make sure the +card ships with a low-profile bracket or buy one separately. + +With a 4-port PCIe card + 2 rear USB 3.0 = **6 USB 3.0 ports** = up to 6 simultaneous drives. + +Each Blu-ray drive reads at ~36 MB/s (1x BD) to ~54 MB/s (UHD). USB 3.0 at 5 Gbps (~400 MB/s +practical) handles many drives per port, but a **powered USB 3.0 hub** is recommended to +ensure each drive gets adequate power. + +### Recommendation: E5-2660 V4 (28 Cores, H730) + +**Winner: E5-2660 V4** — best balance for a media server. + +| Why | Detail | +|-----|--------| +| **28 cores** | Enough for 2-3 simultaneous HandBrake encodes + ripping + Plex | +| **105W TDP** | 60W less than the 2690 = lower power bill, less heat, quieter fans | +| **H730 RAID** | Write cache handles multi-drive rip throughput. Battery-backed | +| **Price** | Should be mid-range of the three listings | + +**Why not E5-2690 V4?** The 3.5 GHz turbo is nice, but encoding is a sustained all-core +workload — turbo barely helps. The 135W TDP means $100+/year more in electricity for maybe +10-15% faster encodes. Not worth it for a media server. + +**Why not E5-2620 V4?** 16 cores is enough for ripping, but if you ever want to rip 3 discs +AND encode simultaneously, you'll feel the pinch. The H330 with no write cache is the real +problem — sustained writes from multiple Blu-ray rips to spinning disks will bottleneck. + +### Before Buying — Ask the Seller + +1. **What PSU wattage?** 750W is fine for no-GPU media server use. Don't overpay for 1100W + you won't need. +2. **"2x trays" — what does that mean?** 2x drive trays in front? Or 2x 2.5" rear flex bay? + You want to know what drive caddies are included. +3. **How many drive caddies total?** The R730xd has 12x 3.5" front bays (LFF model) or + 24x 2.5" front bays (SFF model). Empty bays without caddies = more $ to fill them. +4. **RAM details?** "32GB" could be 2x16GB (6 empty slots per CPU) or 4x8GB or 8x4GB. + Fewer, larger sticks = more room to upgrade later. + +### Accessories You'll Need + +``` +EXTERNAL ODD DRIVES + 2-3x USB 3.0 Blu-ray drive (Pioneer BDR-XD08, LG WP50NB40, etc) ~$60-100 ea + or 5.25" internal drives in USB 3.0 enclosures ~$30-50 ea + +USB EXPANSION + 1x PCIe USB 3.0 card (StarTech PEXUSB3S44V or similar) ~$30-40 + 1x Powered USB 3.0 hub (Plugable USBC-HUB4A or similar) ~$25-35 + +STORAGE + *x 3.5" HDDs for media storage (fill those front bays!) varies + 1x SSD for OS + ripping scratch space (500GB-1TB) ~$30-50 + +SOFTWARE + MakeMKV (free beta / $60 lifetime license) + HandBrake (free, open source) + Plex or Jellyfin (free / Plex Pass $120 lifetime) +``` + +### UHD Blu-ray Ripping Note + +For 4K UHD discs, you need a **UHD-friendly drive** with modified firmware. Not all Blu-ray +drives can read UHD discs — the disc encryption (AACS 2.0) requires specific drive models +that can be flashed. Check the [MakeMKV UHD Drives Guide](https://forum.makemkv.com/forum/viewtopic.php?t=19634) +for compatible models before buying. + ## Sources - [Quadro RTX 8000 for Local LLMs — Hardware Corner](https://www.hardware-corner.net/guides/quadro-rtx-8000-for-llm/) - [RTX 8000 Passive — Network Outlet](https://networkoutlet.com/blogs/articles/nvidia-quadro-rtx-8000-48gb-passive-cooling-powering-ai-rendering-server-workloads) @@ -866,3 +1042,12 @@ INVOKEAI_PRECISION=float16 - [7 Local LLM Families to Replace Claude/Codex](https://agentnativedev.medium.com/7-local-llm-families-to-replace-claude-codex-for-everyday-tasks-25ba74c3635d) - [Qwen2.5-Coder 32B on Ollama](https://ollama.com/library/qwen2.5-coder:32b-instruct-q4_K_M) - [Qwen3-Coder — How to Run Locally](https://unsloth.ai/docs/models/qwen3-coder-how-to-run-locally) +- [MakeMKV Forum: Multiple Drives Ripping](https://forum.makemkv.com/forum/viewtopic.php?f=16&p=173500) +- [MakeMKV Forum: UHD Drives Flashing Guide](https://forum.makemkv.com/forum/viewtopic.php?t=19634) +- [Dell R730xd USB Ports — Owner's Manual](https://www.dell.com/support/manuals/en-us/poweredge-r730xd/r730xd_ompublication/usb-ports?guid=guid-ace748db-1711-4f7c-862d-31ee7600f6e0) +- [Adding USB 3.0 PCIe Card to Dell PowerEdge](https://tachytelic.net/2014/11/dell-poweredge-usb3-pcie/) +- [TrendForce: Q1 2026 Memory Prices Hit Record Highs](https://www.trendforce.com/presscenter/news/20260202-12911.html) +- [Network World: Server Memory Prices Could Double](https://www.networkworld.com/article/4093752/server-memory-prices-could-double-by-2026-as-ai-demand-strains-supply.html) +- [Best Local LLMs for 16GB VRAM (2026)](https://localllm.in/blog/best-local-llms-16gb-vram) +- [Best Local Coding Models Ranked by VRAM Tier (2026)](https://insiderllm.com/guides/best-local-coding-models-2026/) +- [Open Source vs Paid AI for Coding (2026)](https://aarambhdevhub.medium.com/open-source-ai-vs-paid-ai-for-coding-the-ultimate-2026-comparison-guide-ab2ba6813c1d)